And there is this paradox where it becomes harder to detect the problems as the models 'improve'.